[BUG] pytorch not working for video tensors

[verbose-void]

🐛🐛 Bug Report

created a video tensor (mp4), added 35 samples, tried to use the ds.pytorch function, got this error:

⚙️ Environment

ubuntu 20.04, python=3.7.13, jupyter notebook

[verbose-void]

would be nice if i could get a batch of frames instead a batch of videos too

[davidbuniat]

@nollied thanks for reporting the issue, will take a look. @AbhinavTuli the new dataloader we should make sure this works.

[verbose-void]

@davidbuniat cool, btw are the concerns about random shuffling applicable to video datasets?

basically, lets say there's 100 video sequences ~100MB each (so these videos are stored in individual chunks), if my goal is to train a model where each batch has, say, 20 frames but i would like it to be shuffled with respect to the video index (so i don't want to mis-match video frames in any batches), what kind of random sampling guarentees can i expect?